feat(ms-ai-architect): reis idx-33v - modell-parentes uten referent paa kilden, bevisst uten form [skip-docs]

Operatoerbeslutning 2026-08-25 (alternativ A). Funnet ble maalt under
anvendelsen av #34 paa idx-33l (d8e2039) og laa til naa bare i den entryens
resolution - altsaa paa en RESOLVED entry, utenfor aapen-lista og dermed
usynlig for neste sveip. Det er hele grunnen til at den reises.

FUNNET. multimodal-prompt-design.md:196 skriver "GenAI Prompt skill + LLM
(phi-4, gpt-4o, gpt-5)". Den stemplede kilden
(azure/search/multimodal-search-overview) navngir INGEN modell for skillen i
det hele tatt. Skillens egen side (cognitive-search-skill-genai-prompt) sier
"any chat completion inference model deployed in Foundry, such as GPT models,
DeepSeek-R#, Llama-4-Maverick, and Cohere-command-r" - gpt-5 opptrer der bare
i en temperature-begrensning, gpt-4o bare i eksempler, og phi-4 finnes paa
ingen av de to sidene. Begge hentet via microsoft-learn MCP 2026-08-25.

HVORDAN DEN BLE FUNNET, som er den gjenbrukbare halvdelen: blokkstempelet paa
l. 190 dekker 191-198, saa aa flytte det asserterer at hver linje under det
ble sjekket mot kilden. L. 196 navngir modeller den stemplede sida ikke har,
saa skill-sida ble hentet separat FOER stempelet fikk flyttes. Et stempel du
flytter er en paastand du avgir.

BEVISST UTEN FORM. Verken #34 (kildens sett er AAPENT - "any ... such as" -
saa passasjen er ikke en sann paastand som naar kortere enn et determinert
kildesett) eller #35 (passasjen motsier ikke kilden; phi-4 kan godt vaere
deploybar i Foundry, sida sier det bare ikke). Naermeste nabo er idx-33t sin
form, men idx-33t er en droppet kvalifikator og idx-33u er loegnaktig
proveniens paa en sann paastand. Dette er et tredje: en enumerasjon uten
referent paa sida den er stemplet mot. Ett medlem er ingen konvensjon
(idx-33a), saa klassen er latt umaalt med vilje (#29). Ingen gren, ingen
reparasjon foreskrevet.

KONSEKVENS, med leserpopulasjonen oppgitt av operatoeren og foert som
OPERATOEROPPLYSNING, ikke som kildefakta: saa godt som alle norske brukere av
denne pluginen kobler paa modeller via Microsoft Foundry. Det er noeyaktig den
konfigurasjonen kilden beskriver, saa leserens reelle sett er det de har
deployet. En parentes med tre navn leser da som skillens stoettede sett og
styrer dem bort fra en DeepSeek-, Llama- eller Cohere-deployment de allerede
har. Defekten koster leseren valgmuligheter, ikke korrekthet.

Ankeret er den LEVENDE l. 196 (hel-linje, unik, count=1); check-g7-queue
bekrefter at alle 12 aapne fortsatt anker til levende korpustekst.

Ingen KB-fil roert i denne commiten - kun koea. Koea: 79 -> 80 entries
(11 -> 12 aapne, 68 resolved). Diff 12+/0- - ingen reformatering.

Den pendende ordren 20260825T112900Z-3603318296 ble skrevet FOER denne
beslutningen og er derfor sjekket paa nytt og oppdatert: idx-33v er lagt inn
i dens SKAL IKKE ROERES-liste, saa #35-oekten ikke folder den inn fordi den
ligger ved siden av i koea.

Gate: check-g7-queue.mjs 0 drift exit 0; suite 1052/1052.
[skip-docs]: ren koe-endring, ingen kommando-, agent-, skill- eller
hook-flate roert.

scripts/kb-eval/data/g7-review-queue.json

@@ -1294,6 +1294,18 @@
       "anchors": [
         "- [Workbooks programmatic management](https://learn.microsoft.com/en-us/azure/azure-monitor/visualize/workbooks-automate) *(Verified MCP 2026-04)* — ARM/Bicep deployment, RBAC (Monitoring Contributor for redigering, Monitoring Reader for visning), `microsoft.insights/workbooks/write` for custom roles"
       ]
+    },
+    {
+      "id": "idx-33v",
+      "file": "skills/ms-ai-advisor/references/prompt-engineering/multimodal-prompt-design.md",
+      "class": "replacement",
+      "status": "open",
+      "raised": "2026-08-25",
+      "summary": "Raised 2026-08-25 while applying #34 to idx-33l (order 20260825T070602Z-862318346, commit d8e2039). NOT a #34 member, and deliberately given NO form - see the class paragraph below.\n\nHOW IT WAS FOUND, which is the reusable half. The block stamp at line 190 covers lines 191-198, so moving it to the fetch date asserts that every line under it was checked against the stamped source. Line 196 names three models, and the stamped source carries no model names at all, so cognitive-search-skill-genai-prompt was fetched separately before the stamp was allowed to move. A stamp you move is an assertion you make; this finding is what checking that assertion cost.\n\nTHE MEASUREMENT. Line 196 reads \"GenAI Prompt skill + LLM (phi-4, gpt-4o, gpt-5)\". The stamped source (azure/search/multimodal-search-overview, fetched 2026-08-25) names no model for the GenAI Prompt skill anywhere - it says only \"producing concise natural-language descriptions for text search and embedding using a large language model (LLM)\". The skill's own page (azure/search/cognitive-search-skill-genai-prompt, fetched the same day) says under \"Supported models\": \"You can use any chat completion inference model deployed in Foundry, such as GPT models, DeepSeek-R#, Llama-4-Maverick, and Cohere-command-r.\" On that page gpt-5 appears only in a temperature constraint and gpt-4o only in sample definitions and a best-practice line. phi-4 appears on NEITHER page.\n\nWHY IT IS NEITHER #34 NOR #35, stated so it is not folded into one of them by proximity. Not #34: the source's set is OPEN (\"any ... such as\"), so the passage is not a true claim reaching less far than a determinate source set - the same asymmetry idx-33l and idx-33p already turn on. Not #35: the passage does not contradict the source either; phi-4 may well be deployable in Foundry, the page simply does not say so. The nearest neighbour is idx-33t's shape (\"the file names no source for this\"), but idx-33t is a dropped qualifier and idx-33u is a lying provenance on a true claim. This is a third thing: an enumeration with no referent on the page it is stamped against, under a stamp. One member is not a convention (idx-33a), so no form is proposed and the class is left unmeasured on purpose, per #29.\n\nCONSEQUENCE, with the reader population supplied by the operator. OPERATOR STATEMENT 2026-08-25, recorded as such and NOT as a source fact: practically every Norwegian user of this plugin connects models through Microsoft Foundry. That is exactly the configuration the source describes - \"any chat completion inference model deployed in Foundry\" - so the reader's real set is whatever they have deployed. A three-name parenthesis reads to that reader as the set of models the skill supports, and steers them away from a DeepSeek, Llama or Cohere deployment they already have and could use. The defect costs the reader options rather than correctness.\n\nNOT ADJUDICATED, and no repair is prescribed here. Whether the fix is to restate the source's openness, to name the source's own four examples, or to drop the parenthesis entirely, needs its own live fetch and its own verdict. The line is recorded, not disposed of.\n\nREST OF THE BLOCK IS FAITHFUL, checked in the same fetch and recorded in idx-33l's resolution: lines 192-195 and 197-198 all trace to multimodal-search-overview and were confirmed there.",
+      "evidence": "microsoft-learn MCP 2026-08-25 (azure/search/multimodal-search-overview: names no model for the GenAI Prompt skill; azure/search/cognitive-search-skill-genai-prompt, \"Supported models\": \"You can use any chat completion inference model deployed in Foundry, such as GPT models, DeepSeek-R#, Llama-4-Maverick, and Cohere-command-r\" - phi-4 absent from both pages); this file 190 (block stamp, moved to 2026-08-25 in d8e2039), 196; idx-33l (the entry whose application surfaced this, and which carries the same observation in its resolution); idx-33t and idx-33u (the two nearest neighbours, neither of which is this shape); _meta.ratified_forms[\"#34\"] and [\"#35\"] (both checked and neither governs)",
+      "anchors": [
+        "3. **Image verbalization:** GenAI Prompt skill + LLM (phi-4, gpt-4o, gpt-5) → naturlig-språklig beskrivelse"
+      ]
     }
   ]
 }